FACE-GRAB: Face recognition with General Region Assigned to Binary operator

Face recognition in unconstrained environments is one of the most challenging problems in biometrics. One vexing problem in unconstrained environments is that of scale; a face captured at large distances is considerably harder to recognize than the same face at small distances. Several methods have been proposed to tackle unconstrained face recognition in a robust fashion, including face descriptors such as LBP and its many variations, of which some are less sensitive to scale variations than others. In this paper, we present a novel operator called General Region Assigned to Binary (GRAB), developed as a generalization of LBP. We demonstrate its performance for face recognition in both constrained and unconstrained environments and across multiple scales. Unlike prior work, the GRAB descriptor accounts for multiple scales and resolutions through the size and choice of its neighborhood and is evaluated with respect to varying scales. We show that GRAB significantly outperforms LBP in cases of reduced scale on subsets of two well-known published datasets of FERET and LFW, introducing useful subsets of these datasets for recognition system evaluation.